MiniMax releases MiniMax-M1, an open reasoning model with 1M context

Background

MiniMax is a Chinese AI company that developed the abab line of proprietary text models. In January 2025 it released the open-weight MiniMax-01 series, a language model and a vision-language model ^[2][3] M1 extended that open line into reasoning.

Reasoning models are trained to work through problems step by step before answering. DeepSeek R1 is one of the reasoning models the release figures use for comparison.^[1]

What happened

In June 2025 MiniMax released M1, a reasoning model.^[2] It is a hybrid mixture-of-experts model with lightning attention, with 456 billion total parameters and 45.9 billion activated for each token.^[1] Mixture-of-experts means that only some specialist sub-networks run on each input, so the active part of the model is far smaller than its total size.

The model natively supports a 1-million-token context, meaning it can consider up to a million tokens of text at once.^[1] A token is a unit of text such as a word or part of a word. In practical terms, a context that long lets a model take in very long documents within a single request.

Key details

At a generation length of 100K tokens, M1 uses 25% of the FLOPs that DeepSeek R1 uses, according to the figures published with the model.^[1] FLOPs measure the arithmetic a model performs, so this figure is a measure of computational efficiency during long reasoning.

M1 was trained with CISPO, a reinforcement-learning algorithm that clips importance-sampling weights rather than token updates.^[1] The release is open under the Apache 2.0 license.^[1]

Impact

The release gave outside researchers an open reasoning model with a 1-million-token context. It showed that a long-context architecture could be applied to step-by-step reasoning. The release figures compared M1 with DeepSeek R1 on efficiency.^[1]

For the archive, the event marks MiniMax's entry into open reasoning models, which were followed by M2 in October 2025.^[3]